Many papers describe classifiers of speckle pattern images obtained as a result of interference of light going through quasi-monomode optical fiber. The feature extraction is achieved by placing in Fourier plane a computer generated hologram (CGH) which serves as ring wedge-detector (RWD). The basic advantage of using CGH instead of RWD is its potential possibility to be easily changed, and thus, optimized for current classifications. This paper presents a new method based on rough sets theory (RST) and evolutionary algorithms, aimed to obtain the optimal CGH-based feature extractor for given classification problem. The task of CGH is dimensionality reduction of different pattern, while preserving all features necessary for further classification. The goal of optimizing feature extractor in terms of RST is to find such set of conditional attributes, for which approximation quality with respect to decision attribute has maximum value. Since there is no gradient direction information involved in above indicator, use of it as an objective function, depends stochastic method, such as evolutionary optimization. In the end, neural network fed with features extracted by CGH is presented, as the experimental confirmation of good classification abilities of the whole system.
|